Skip to content

Upgrading from Upstream LiteRT

heliaRT is a drop-in replacement for upstream LiteRT for Micro, formerly TensorFlow Lite for Microcontrollers / TFLM. The API surface is identical, so switching requires only changing where the library comes from.

What Changes

Aspect Before (upstream LiteRT) After (heliaRT)
Source / archive tensorflow/tflite-micro AmbiqAI/helia-rt
Additional backend Reference + CMSIS-NN Reference + CMSIS-NN + HELIA
Build variants Single SPEED and SIZE
Toolchain support GCC, armclang GCC, armclang, ATfE
Ambiq-specific tuning None heliaCORE kernel paths

What Stays the Same

  • .tflite model format — no retraining or re-quantization needed
  • MicroInterpreter lifecycle (AllocateTensors → Invoke → read output)
  • MicroMutableOpResolver registration pattern
  • Tensor arena sizing and static memory planning
  • All upstream Reference and CMSIS-NN kernels remain available

Step-by-Step: Zephyr

Replace the upstream LiteRT module with heliaRT in your west.yml:

# west.yml — projects:
- name: helia-rt
  url: https://github.com/AmbiqAI/helia-rt
  revision: main        # or pin to a release tag
  path: modules/lib/helia-rt

Update prj.conf:

CONFIG_HELIA_RT=y
CONFIG_HELIA_RT_BACKEND_HELIA=y   # or CMSIS_NN / REFERENCE

Download the latest release archive from GitHub Releases and point your west.yml at the extracted directory. See the Zephyr getting-started guide for details.

Step-by-Step: Makefile / Source Build

  1. Clone heliaRT instead of upstream LiteRT:

    git clone https://github.com/AmbiqAI/helia-rt
cd helia-rt

  2. Build with the HELIA backend:

    make -f tensorflow/lite/micro/tools/make/Makefile \
    TARGET=cortex_m_generic \
    TARGET_ARCH=cortex-m55 \
    OPTIMIZED_KERNEL_DIR=helia \
    microlite

  3. Link the resulting libtensorflow-microlite.a into your project exactly as before.

Step-by-Step: neuralSPOT

neuralSPOT already bundles heliaRT. If you're using ns_autodeploy, you're already on heliaRT — no migration needed.

Verifying the Upgrade

After building, confirm heliaRT is active:

#include "tensorflow/lite/micro/heliart_version.h"
printf("heliaRT %s\n", HELIART_VERSION);

Next Steps